Product Defect Detection Based on Transfer Learning of CNN

カテゴリ: 研究会(論文単位)

論文No: ST18099

グループ名: 【C】電子・情報・システム部門 システム研究会

発行日: 2018/09/27

タイトル(英語): Product Defect Detection Based on Transfer Learning of CNN

著者名: Su Kai(University of Aizu),Zhao Qiangfu(University of Aizu)

著者名(英語): Kai Su(University of Aizu),Qiangfu Zhao(University of Aizu)

キーワード: Product defect detection|Convolution neural network (CNN)|Image classification|Transfer learning|Product defect detection|Convolution neural network (CNN)|Image classification|Transfer learning

要約(日本語): In this paper, we focus on product surface defect detection. Defect detection is an essential step in a production line. So far, this task has been conducted mainly by human inspectors. The inspection results are often affected by various human factors like inspector’s experiences, health conditions, and so on. To improve the accuracy, in this study we apply the convolution neural network (CNN) to support the human inspector. In recent years, CNN has been applied successfully for image recognition in various fields. In this paper, we investigate several methods based on CNN, and report results obtained through experiments on image datasets provided by our partner company. Results show that both AlexNet and GoogLeNet can recognize surface defect very well with the recognition rates 99.63% and 99.51%, respectively. The proposed system can “reject” a certain percentage of the data and leave them for human-based inspection. In addition, the system can also detect wrongly labeled data or outliers, and thus can help human inspectors to purify the training data.

要約(英語): In this paper, we focus on product surface defect detection. Defect detection is an essential step in a production line. So far, this task has been conducted mainly by human inspectors. The inspection results are often affected by various human factors like inspector’s experiences, health conditions, and so on. To improve the accuracy, in this study we apply the convolution neural network (CNN) to support the human inspector. In recent years, CNN has been applied successfully for image recognition in various fields. In this paper, we investigate several methods based on CNN, and report results obtained through experiments on image datasets provided by our partner company. Results show that both AlexNet and GoogLeNet can recognize surface defect very well with the recognition rates 99.63% and 99.51%, respectively. The proposed system can “reject” a certain percentage of the data and leave them for human-based inspection. In addition, the system can also detect wrongly labeled data or outliers, and thus can help human inspectors to purify the training data.

原稿種別: 英語

PDFファイルサイズ: 1,737 Kバイト